稀疏地理实体关系的关键词提取方法

网络文本蕴含地理实体关系抽取技术,需要高时效、强鲁棒的关键词提取方法。与监督学习方法相比,无监督学习方法能捕获文本的动态变化特征并发现新增的关系类型,因此备受关注。其中,基于频率的关键词提取方法获得广泛研究,然而,网络文本蕴含的地理实体关系分布稀疏,基于频率的方法难以直接应用于地理实体关系的关键词提取。为解决该问题,本文基于公开访问的网络资源,提出一种语境增强的关键词提取方法。首先,基于在线百科和开放的同义词词典,通过语境合并和语义融合创建增强的语境,以降低语境中词语的稀疏性。接着,Domain Frequency和Entropy频率统计方法从增强语境中自动构建一个大规模语料。然后,基于该语料选择词法特征并统计其权值,用于扩大语境中词语间的差异。最后,使用选择的词法特征度量增强语境中词语的重要性,将权值最大的词语作为描述地理实体关系的关键词,并基于大规模真实网络文本开展实验。实验结果表明：对于地理实体关系的关键词识别,本文方法的平均精度为85.5%,比Domain Frequency和Entropy方法分别提高41%和36%;对于新增关键词识别,本文方法的精度达到60.3%。语境增强的关键词提取方法能有效地处理地理实体关系分布的稀疏性,可服务于网络文本蕴含地理实体关系的抽取。

A Method of Context Enhanced Keyword Extraction for Sparse Geo-entity Relation

Geo-entity relation recognition from rich web texts requires robust and effective keyword extraction method. Unsupervised learning methods attract more attention because they can capture dynamic variations of features in text and discover additional relation types. Frequency-based methods for keyword extraction have been extensively studied. However, the sparse distribution of geo-entity relations in web texts makes it difficult to directly apply frequency-based methods to geo-entity keyword extraction. This paper proposes a context enhanced keyword extraction method to solve this problem. Firstly, the contexts of geo-entities are enhanced to reduce the sparseness of terms, with context merging and semantic fusion. Secondly, two well-known frequency-based statistical methods (Domain Frequency and Entropy) are used to automatically build a large-scale corpus. Thirdly, the lexical features and their weights are statistically determined based on the corpus. Finally, all terms in the enhanced contexts are measured according to their lexical features and the most important terms are picked as keywords of geo-entity pairs. Experiments are conducted with large and real web texts. The results show that compared with the Document Frequency and Entropy methods, the presented method improved the precision by 41% and 36%, respectively. It also correctly generated additional 60% of keywords.